<h1 id="krea2-finalcut-fusion">Krea2 FinalCut Fusion</h1>AMD, Nvidia, Intel, MacIf your planing on using low resolution 832x1216 or don't need the line art/anime fusion training use <a target="_blank" rel="ugc" href="https://civarchive.com/models/2811049/krea2-finalcut-unique-138gb-int8-convrot">Final Cut Base</a> instead of this model.This model defaults to nude if clothing is not prompted.<h3 id="fp16">FP16</h3>This version is recommended it is more flexible then FP8 or comfy quant however if the seed or prompt decides on ahego face or "wet skin" it is hard to avoidCountless hours of work and training have gone into this custom quantization of Krea2. This version has more training then V2 of FinalCut but is slightly less photorealistic do to the anime and other art-style training.<ul><li>Line Art Style</li><li>Anime</li><li>Realistic</li><li>4-Steps is possible with SGM_Uniform and Euler_A</li><li>8-Steps with any Turbo Compatible (Euler, DPM2, Etc) up to 2048x2048</li></ul><hr /><h3 id="custom-comfy-version:">Custom Comfy Version:</h3><a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3268260?fileId=3152072">INT8 Conrvot</a>: All those using comfy should likely use the INT8 Conrvot version. With a few exceptions.MXFP8: 50xx users who benefit from MXFP8 increase in speed and use full FP8 calculations (Not cast to BF16)Scaled FP8: 40xx may see a slight increase in speed compared to INT8 if using full FP8 (Not cast to BF16)

Krea2 FinalCut Fusion (FP16, INT8, MXFP8, ScaledFP8)

krea2_finalcut_pro_base_fp8.safetensors

qwen_image_hdr_vae.safetensors

Universal_FP8

Universal_FP16

Comfy_Quants

Note: all models not trained on NSFW, you will need to use other lora or checkpoint with my loraEnhances the textures and lighting of krea. I used the int8 of the krea model. I also recommend using @AiMetatron redcraft (Krea2RedMix1.1-INT8-Convrot-ComfyUI) model (can also be found in my posts). This lora adds extra noise that would benefit that model. Using the original krea model may be noisy, though lowering the strength may help and still usable i think. Its still experimental. Im still using my own custom nodes.Version 2 fixed: Made some improvements to the lora. I also did some changes to my custom nodes and i think i cooked with them. My results will be different than yours. This version will work great for the original krea2 model from comfyui. Support me by uploading your results! The more people that use my lora and post their results, the higher the chance i release my workflow. have Fun!I highly suggest using @Felldude vae model: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2718533/hdr-vae-anima-qwen-image">https://civarchive.com/models/2718533/hdr-vae-anima-qwen-image</a>I know it says its for anima, but it also works for krea2 if you are wondering. It increases vibrant and contrast of your images, doing it with just the right amount. All my images were used with the hdr vae model.Version 6 ultimate : this is a special lora that is capable of doing almost anything as long as its prompted correctly. It handles both realism and anime/art styles. It just simply enhances the krea 2 model.Edit: For the ultimate version, i just uploaded a more stable variant of the lora. Ultimate ver might be a bit overcooked at 1.00 strength. Stable ver fixes that

Mrpopo's Krea2 photoreal enhanced

Did some changes to the dataset and experimental tests. This took a long time to get it to how I wanted it to be. Though, still not perfect, I find it performing better than the last model. I'll upload more images soon. I recommend using 1.50 strength. Try 1.00 first. I'll leave it as 100 buzz to access it for now. Then make it available to everyone within a couple days. If you like my work please show support. A lot of time has gone through making the loras. Thank you

hdrVAEAnimaQWENImage_fp32.safetensors

Mrpopo new dataset realism.safetensors

Aesthetic realism v5.0

Mrpopo's Real Clarity

Krea Trained High Res 6.0

Mrpopo photoreal ultimate

realism fixed v4.0

improved realism v3.0

Fixed versionV2

v1.0

For V3.0: i recommend using 10 stepsI did a thing. Experimental, but so far good. This lora enhances the color and detail of krea2 for anime. Enjoy!Note: i am using my own custom nodes such as scheduler, sampler, and some other important node. My results may be different than yours. However, without my nodes, it still produces great results nonetheless. In case people were wonderingI highly suggest using @Felldude vae model: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2718533/hdr-vae-anima-qwen-image">https://civarchive.com/models/2718533/hdr-vae-anima-qwen-image</a>I know it says its for anima, but it also works for krea2 if you are wondering. It increases vibrant and contrast of your images, doing it with just the right amount. All my images were used with the hdr vae model.

Mrpopo's Krea2 Anime/Art enhanced

I've been working on this lora for a while now. I made a new lora, which that alone was really good and was supposed to be the next version, but i decided to experiment with some new techniques. The result... Insane. It is exactly the art style that i've been trying to achieve. It could be used with various checkpoint models (i've tried only 4 so far). Just try it for yourselves. As always please upload your results, i would love to see them.

Mrpopo anime full potential.safetensors

Full Potential v3.0

v2.0 improved the lora

Very big improvements. There has been one thing that bugged me with krea2. The fact that it tends to smooth out the skin or even look very grainy. So i decided to fix that. It required a lot of testing and cherry-picking for the right images for the dataset. With the help of claude fable, being my master prompt maker, i was able to get the dataset that i wanted to truly fix the annoying problem in krea2. I've always new krea2 had the potential of being more than it is now. My lora is the key that krea needs to go beyond. This is just the start, there is always room for improvements. As always, please show support by uploading your images in my model post. Have fun!

photoreal_epoch_10e20.safetensors

I made a mistake. I accidentally used and uploaded the wrong lora. The v3.0 is my overcooked version which is not meant to be used (unless you lower strength, but thats not what i wanted). This new version is the correct version and will give expected results at strength 1.00.

photoreal_epoch_10e9.safetensors

i've been working on this for a while now and im very pleased with the results It is better than all the previous versions guaranteed. The lora was trained on 2048 res so expect insane level of detail. There was some complications when training at that particular res. I finally figured it out and now got this lora. It was pretty expensive (used a cloud gpu) but worth it. Note: the only upscale i used is just rtx super resolution node. The details are from the lora.

krea realism.safetensors

This one is a full retrain, not a merge. I went back through the whole dataset and rebuilt it properly before training a fresh LoRAIt's better than the last lora across the board (in my opinion). Because the dataset is so much cleaner, it follows your prompts better, and it can handle almost any art style you throw at it. Detail is sharper too, since it was trained on higher resolution images.My other anime lora is still good as it is, just different style compared to this lora. just a matter of taste.What changed:- Every caption was checked against its image. Anything the prompt described that wasn't actually in the picture got fixed, and images with no caption got one written.- Higher quality training images- Cleaned up the files themselves (transparent PNGs and .webp converted to plain RGB PNG).<ul><li>Recommended settings:</li></ul>- 10 steps, CFG 1.0- No trigger word needed- I still recommend @Felldude's HDR VAE, it works great with this.As always, please upload your results, I'd love to see them.

Mrpopo's krea2 Anime hi res

Mrpopo's krea2 Anime hi res.safetensors

I am pleased to release this model. It is the cultivation of all the previous models before this. In my opinion, it may be my finest work compared to all my loras. I can't wait to see what you guys can generate with it. I'd like to thank the people who showed support in making this possible (i used up most of my buzz to train the loras, of which a lot of them failed expectations). you could download it using either yellow or blue buzz. Just click the download button to show option of blue buzz.

Mrpopo photoreal ultimate edition.safetensors

krea stable non merge version.safetensors

this might be the limit to what i can do to make it look "realistic". The details are more tamed than the last version. the prompt adherence is a bit better. the lora has better compatibility with other loras, improved compared to last version. you dont even want to know how much it costed me to finally get this result. i dont even want to think about it lol. Anyway, have fun with this lora! :) -used ai-toolkit using b200 gpu. Note: this is best used with higher resolutions.

Mrpopo's Real Clarity.safetensors

<h1 id="qwen-image-vae">Qwen Image VAE</h1><ul><li>Full FP32 Training of Decoder</li><li>Works in ComfyUI</li></ul><hr /><h2 id="overview">Overview</h2>This model is a fine-tuned variant of the base Qwen Image VAE, modified to emphasize high-frequency detail preservation and expanded color representation, following an HDR-style reconstruction objective.The evaluation compares the base and HDR-tuned models using perceptual, structural, distributional, and photometric metrics over identical input data.<hr /><h2 id="evaluation-summary">Evaluation Summary</h2><h3 id="perceptual-fidelity-(lpips)">Perceptual Fidelity (LPIPS)</h3><ul><li>Base: 0.0177</li><li>HDR: 0.0786</li></ul>The HDR model exhibits a significant increase in perceptual distance, indicating reduced strict identity reconstruction under deep feature similarity metrics and a shift toward detail-enhancing reconstruction behavior.<hr /><h3 id="structural-energy-(gradient-magnitude)">Structural Energy (Gradient Magnitude)</h3><ul><li>Ground Truth: 404.02 (both models)</li><li>Base Reconstruction: 313.46</li><li>HDR Reconstruction: 687.97</li></ul>The base model demonstrates strong low-pass behavior with reduced high-frequency content. In contrast, the HDR model exhibits high-frequency amplification, exceeding the structural energy of the original inputs.<hr /><h3 id="color-distribution-support">Color Distribution Support</h3><ul><li>Ground Truth: 33150.61 (both models)</li><li>Base Reconstruction: 35004.49</li><li>HDR Reconstruction: 40133.37</li></ul>The HDR model produces a substantially expanded color support space, indicating increased chromatic dispersion and reduced quantization collapse.<hr /><h3 id="photometric-stability">Photometric Stability</h3>Brightness Bias<ul><li>Base: 0.000351</li><li>HDR: 0.0000098</li></ul>Contrast Gain<ul><li>Base: 0.9984</li><li>HDR: 0.99999</li></ul>Both models preserve global photometric consistency, with the HDR variant showing near-perfect affine stability.<hr /><h3 id="channel-drift">Channel Drift</h3><ul><li>Red Shift:<ul><li>Base: +0.0116</li><li>HDR: +0.0104</li></ul></li><li>Green Shift:<ul><li>Base: -0.0606</li><li>HDR: -0.1856</li></ul></li><li>Blue Shift:<ul><li>Base: +0.0187</li><li>HDR: +0.0219</li></ul></li></ul>The HDR model introduces a significantly stronger negative bias in the green channel, while maintaining comparable red and blue stability.<hr /><h2 id="interpretation">Interpretation</h2>The base Qwen VAE behaves as a contractive perceptual projection operator, prioritizing smooth reconstructions and suppression of high-frequency components.The HDR-tuned variant transitions into a detail-amplifying reconstruction operator, characterized by:<ul><li>Increased high-frequency energy</li><li>Expanded color manifold coverage</li><li>Higher perceptual divergence under LPIPS</li><li>Preserved global photometric invariance</li></ul>This represents a functional shift from a smoothing autoencoder regime toward a high-frequency preserving (HDR-like) reconstruction regime.

Files with SHA256: c3c0d4f85dd3e70d64685febf55df7d49e9f8473bfdd8a769229787be3aa078f